Add unit tests for kernel implementations

[lockshaw]

While operators are currently tested with alignment tests, these test the full FlexFlow stack (python -> c++ -> cuda), which makes bug localization difficult. Adding unit tests for the raw kernel implementations would improve stability/correctness, increase readability (by making clear which pointers should be non-null, which are fine to be null), and improve error localization.

May depends on non-parallel tensor shape inference from #948 depending on how the tests are written

[lockshaw]

For now, this should focus on just exercising the test code and making sure it doesn't crash--the story for testing functionality of the kernel code (just alignment tests, additional C++ tests, etc.) will be decided later